Scaling AI Transformations... with AI!

You can refer to the original version of this article that I wrote recently in Forbes.

According to a recent McKinsey survey, just under a third of organizations use artificial intelligence for multiple business functions. Scaling AI transformations appears to be the most important challenge in this arena, and based on my experience, the key limiting factor here is that AI projects are typically human-workload intensive, resulting in lengthy and expensive projects.

But what if AI could help organizations implementing AI?

New technologies and concepts have recently come to the market to help accelerate and improve the AI implementation process. While most of these technologies are still maturing, they have already delivered significant benefits to the organizations that have adopted them.

AI implementation projects typically include:

1. The identification and assessment of AI opportunities.

2. The design and implementation (including coding) of the AI programs.

3. The maintenance of these AI programs.

For each of these three steps, I will describe the new concepts available and their impacts.

1. Identification And Assessment Of AI Opportunities

Selecting the appropriate AI opportunity to implement is critical. Nevertheless, process and data analysis, documentation, assessment and prioritization are workload-intensive. They consist of interviewing, observing, collecting and analyzing data. As a result, this phase often requires two to six months of resource-intensive work.

AI opportunities can be identified at two levels: process or data. At the process level, two technologies are available: process discovery and process mining. At the data level, the technology is referred to as data discovery.

1.1 Process Discovery

The first process discovery technology was launched in June 2018 by Kryon Systems. Here are the key steps it uses:

1. Observation: A program is installed on the users' computers. While users are performing their day-to-day work, it seamlessly records their clicks, user interface objects and their process steps, and it takes screenshots. This data is sent to a machine learning application for analysis.

2. Process Assessment: After a few days of recording, you're left with a dashboard that presents a list of the processes that were observed. The system ranks them by their potential benefits of being automated, analyzing criteria such as the length of the process or the number of people performing the process.

3. Detailed Process Analysis: The dashboard presented should let you access documentation for each process, which comes in the form of flowcharts that show process variants.

Here is a short informative demo of Kryon Process Discovery:

In my experience, this type of solution can help accelerate AI implementations three to five times faster than normal while increasing the number of use cases discovered by about two.

1.2. Process Mining

Launched by startup Celonis in 2016, modern process mining solutions serve the same objectives as process discovery tools do. Their difference lies in the way they analyze the process data. As opposed to process discovery solutions, which use computer vision and user-interface object recording, process mining solutions use the logs extracted from systems like ERPs.

Here is a short explanation of process mining by Celonis:

Process mining and process discovery solutions can be used in conjunction to improve an outcome. Process discovery is usually less accurate, but it offers a more comprehensive view of the potential across all processes. In contrast, process mining provides the precise detail of each process execution but only on the systems generating structured logs. Processes performed on other applications like Excel, email or PowerPoint cannot be recorded.

1.3. Data Discovery

Finding relationships between data that can drive business value consumes resources and time. Instead of manually testing a hypothetical outcome against a dataset, data discovery solutions scan massive amounts of data to discover thousands of hidden drivers behind strategic business challenges. These solutions also combine companies' information with external sources (e.g., economy, weather, demographics) to reveal hidden patterns and deeper insights.

For example, a data discovery solution was implemented with a global payment company. In just five weeks, it improved fraud detection by 7% with cost savings of $140 million. 

Here is the impressive demo of Spark Beyond data discovery technology:

2. Design And Coding Of AI Programs

2.1. Automation Code Generation

Technology vendors have started to create programs that are able to generate robotic process automation code directly by using the outcome from process discovery or mining solutions. What is so exciting about these programs is that they automatically create and add automation workflows directly into the automation design studio. Developers can then further refine the code. Based on my experience, about 60% to 70% of the code for most AI projects can be pre-generated, doubling the speed of implementation.

2.2. Automated Machine Learning (AutoML)

While data discovery platforms help data scientists create value by identifying relations between data, AutoML solutions support data scientists building their models.

In a typical machine learning application, data scientists have a dataset consisting of input data points for training. Typically, the raw data is not in a suitable format that could be fed into algorithms. Rather, a data scientist has to apply methods of data pre-processing, features engineering and selection that make the dataset suitable for machine learning applications. After these pre-processing steps, data scientists then select algorithms and optimize their parameters to maximize the predictive performance of their machine learning model. Each of these steps has its challenges and involves significant time and resources. AutoML systems help automate these steps.

Watch this amazing demo of Datarobot AutoML:

3. Autonomous Maintenance Of AI Programs

When organizations deal with hundreds of AI programs, managing the changes and failures is challenging. When organizations combine different technologies to automate end-to-end processes with AI, the failure of any of these components often causes the entire process to fail.

One effective way to mitigate this issue is the use of a system that predicts and identifies the changes in the program's environment. Such systems are able to proactively adjust the environment (if the change is due to an environment failure) or the automation program (if the program needs to be adjusted). In case the change cannot be performed automatically by the system, it alerts a person to address the issue. Startup Choiceworx has created such an innovative automated platform.

To get started, meet with your AI implementation team to identify where they spend most of their workload or have the most pain points. These are certainly the areas where you can generate the maximum benefits from using the above levers.

Thanks for reading this article. Please share your views and experience?

For more insights, read the first reference book on Intelligent Automation and Hyperautomation, to be released on October 14th 2020: www.intelligentautomationbook.com

Note: The views reflected in this article are the views of the author(s) and do not necessarily reflect the views of any company or organization

#innovation #technology #intelligentautomation #machinelearning #digitaltransformation #analytics #hyperautomation #artificialintelligence #datascience #futureofwork #robot  #ai #roboticprocessautomation #chatbot #pascalbornet